Veedrac

Y'all wouldn't believe this but 10 years ago AGI was a hundred years away.

> Apollo for comparison took about 8 years from announcement to moon landing. Space shuttle took about 9 years from approval to first flight. Starship has been in development now about 9 years.

This comparison is very unfair. Nine years ago the big rocket was a dream, not even Starship at the time.

> During his presentation, Musk joked that his strategy for raising money might be to “steal underpants,” do a Kickstarter campaign … and profit.

Contra Saturn V, which had strong funding out the gate.

The hard part of reuse is getting the thing back in flying shape. The booster part is actually just demonstrated, and this flight got us a fully-intact upper stage back, sans a few tiles lost to the wind and the landing being simulated.

I can see room for skepticism on their rapid reuse plans, but skepticism on practicable reuse alone just seems discordant with the demonstrated success.

Very clean flight, almost all the way through, despite the intentional missing tiles and new flight pattern. No flap burn through, no issues with simulated Starlink deploy, I don't think they even lost any engines.

They're clearly almost ready to scale this thing, if the next block version doesn't add a ton of problems back on. I'm not sure they're quite at the point of rapid reuse looking feasible, since tiles did come loose near the end of flight; not a problem for stage return, but definitely bad enough to warrant a meaningful correction before a (counterfactual) reflight.

Overall they've clearly proven the recipe works.

This is quite loosely stated. It's true Boston Dynamics is an old company and that they've had some very cool demos. It's not at all true that they've been showing qualitatively similar things for 20 years.

The oldest video on their YouTube channel is 16 years old, and is of a quadrupedal robot not falling over while inching along tricky surfaces.

There are a few exacerbating factors.

The first is that IIUC, CFCs release chlorine atoms which catalyze ozone, whereas aluminium oxide catalyzes the creation of chlorine atoms from chlorine reservoirs, which then go on to catalyze ozone. I loosely believe at this point after some sketchy research and maths that this makes it around two orders of magnitude more potent.

The second is that these particles are produced directly in the upper atmosphere. I couldn't give you a number for how much that changes things, but I assume it's nontrivial.

The final point I've noticed is that mass to orbit has been increasing at a rapid exponential rate recently, and it would not surprise me at all to soon see an extra order of magnitude on it.

Worst-case, that could change your 5,000 year figure to just a couple. I don't think it's that bad, I'm not overly concerned about this issue, but given ozone depletion is a legitimate existential threat and the numbers don't immediately make it seem impossible, I think it's worth paying attention to.

This is correct from the perspective of direct health hazards, but there are still plausible risks. We know from history you don't need a lot of mass to cause global problems, if the material is catalytic.

The intro to the article explicitly states that it will put aside any fanciful ideas that the technology might be better in the future; that "rather than try to forecast the future as it might turn out, I’d prefer to describe reality as it already exists".

The article acknowledges that AI progress to date has worked. It snidely disagrees without argument that AI the field could work from here on out.

I was thinking more wrt. the top end, which did stall out for 30 years from WWII but has otherwise been on fairly smooth exponentials since the beginning.

For sure; the US kills literally hundreds of thousands of people in ways other countries have solved, and bigger faster vehicles seems at odds with the lack of driver and infrastructure responsibility here. I don't want to make light of that.

I just had the numbers run to check this. About 650,000 fewer people would have died over my short life so far, if the US had the vehicle fatality rate of my home country.

It's wild that after a hundred years there is still exponential progress in the power output of cars. The most unusual part to me is how EVs are fundamentally a consumer technology, so it all rapidly falls into mass production territory; eg. Xiaomi sells a 1527hp car for $73k. Horsepower is rapidly reaching 'solved' territory; even at its record speed, BYD's car wasn't even power limited.

A refusal to even acknowledge that AI might work isn't a very sensible refutation of the risks we're going to face.

Because clearly what Acemoglu did was look at current AI systems, noted that they are pretty good at providing contextually relevant information to people, dumped it in the ‘provides contextually relevant information to people’ economic bucket, and then painted a future of _what if that, but slightly more people were involved?_ in the context of their current pet political peeve.

> Think of AI providing useful, context-specific information to electricians, blue-collar workers, nurses, educators and others,

This prediction is insane. Pong was invented 53 years ago. The Apple II isn't even 50 years old. Nor is Space Invaders.

I'm perennially reminded of the Slate Star Codex quote,

> Madden was an Anglican clergyman in 18th-century Ireland, and maybe the first futurist. In 1733, he published Memoirs of the Twentieth Century, a novel about people in 1999 sending letters back through time to tell their 18th-century predecessors what the future would hold.

> How did the prognosticators of 1733 imagine the future? Was it utopian? Decadent? Miserable? Beautiful? Incomprehensible?

> Actually, it was none of those things. It was exactly like 1733 in every way, and the future people were just writing back to remind everyone how much Catholics sucked.

https://slatestarcodex.com/2017/10/09/in-favor-of-futurism-b...

Whatever your prediction of the future, let it be less bad than that.

Nobody knows.

Flight 11 is the last of the 'Block 2' Starships, so not 11, but potentially Flight 12 could be a successful derisking mission of Block 3, and then Flight 13 would be first full operational mission. Actual full reuse could follow soon after that.

The optimistic case is that they've demonstrated everything they need to except for having a fully reliable heat shield, but they also seem to have a workable plan for the heat shield for literally the first time in the program's history.

The pessimistic case is that Block 2 was rough, and making mistakes doesn't prevent an organization from making more mistakes. If things continue to go wrong, it could be arbitrarily far out.

Wikipedia has a compilation of what's on the agenda:

https://en.wikipedia.org/wiki/List_of_Starship_launches#Futu...

I don't think there's a simple answer to this. It's a lot of things all at once.

It's pretty obvious that the programme has had issues that I don't think SpaceX were prepared for, in the parts that aren't the a priori obvious hard problems. In that sense, it just hasn't been as well executed as they'd like. Especially their difficulties with weight and payload that are probably why they frontloaded so many design changes before letting it settle.

Then there's some significant fraction where SpaceX really has intentionally done the 'rapid iteration' thing, and keeps flying unfinalized designs in boundary-testing ways on an in-progress factory system. For all Starship has flown badly, it's not like even the easier competition is flying much well in the meantime. Starship flies absurdly frequently.

And this is in part that SpaceX aren't just building a rocket with 2-3x the liftoff thrust of a Saturn V, aren't just building a rocket with first stage reuse, aren't just building a rocket with unprecedented complete second stage reuse, they're also building it to be cheap and mass produced in a factory. Big is hard. Reuse is hard. Cheap is hard. Factories are hard. Makes sense the rocket is hard!

Then there's also the point that while Starship has been a mixed bag, it has also achieved a lot. Its last flight was extremely solid, and really did show the whole mission profile down to soft and positionally-accurate drill landings of both stages. Nobody else has done what Starship has done.

I frequently find myself hitting repeat. I have over a dozen I've listened to, according to the stats, more than 50 times each. By your definition that's interesting. QED?

This seems broadly correct? Industrialization was amazing for people's standard of living, but it absolutely meant that the average physical good became detached from their craftsmen's learned and aesthetic preferences.

I've been avoiding trying to class what counts as 'interesting' because it feels like the wrong point to make, for a medium so entwined with personal preferences that are so often at odds with each other. To single out my tastes feels like it would be saying 'paleontology is the true interesting subject, unlike architecture'. The tracks you listed here aren't really my thing, but I can see why AI audio doesn't help with it. That's fine.

I will also repeat that I'm well aware that the best stuff is definitely all human. It's not my genre either, but traditional composers like Bach certainly made extremely interesting, clever, even deeply-studiable pieces and AI 'in the style of' those composers surely won't capture much of that. There's a lot of stuff AI can't do wholesale; one particularly strong example is if you're Jacob Collier, AI is not going to make the complex harmonizations and song structures there.

AI is pretty bad at these textural or instrument exploration things like from Collier above or Mike Dawes or Yosi Horikawa or Yoko Kanno or Keiichi Okabe. There's a bunch of music I listen to because it's generically a genre or mood I like and it's well produced, which I won't list here, and AI audio can often do stuff like that at baseline but not especially well. There's also nostalgia; I'm also certain a huge part of the reason I like the Celeste soundtrack so much is in part that I liked the game so much.

But then there's a whole category of music I listen to where the texture is supplemental to the part that defines it, like most of Acapella Science or Bug Hunter or Tom Lehrer. Eg. Prisencolinensinainciusol isn't interesting to me because it's musically complex; the part I care about is that it's a listenable execution of an idea, not precisely how it was executed on. I don't keep coming back to I Will Derive by some random schoolkids recorded on a potato 17 years ago annually because it's sung well or they were particularly clever with how they took another song and changed the words; I come back to it because it's fun and reflects for me onto a part of my past that I remember fondly, and these things make me happy.

All these words and I've still only addressed half the comment. Ok, let's consider the idea that it's not enough for AI audio to facilitate the creation of interesting musical pieces, and it instead has to create whole interesting musical styles. I take issue with this in a bunch of places. I don't reject artists who I judge not likely able to create a new Bossa Nova. I judge artists based on whether the output they produce is something I want. I do the same for AI.

I also think the question about whether AI could 'create' a new style is somewhat misplaced. A style is a cultural centroid, not just a piece of audio. AI can definitely create new musical textures or motifs, but it's always being pulled towards the form of what it's being asked to produce. As long as we're talking about systems that work like today's systems, the question still needs to involve the people that are selecting for the outputs they want. Could that connected system create something as distinct and audibly novel as a new genre? Yeah, probably, given time and a chance for things to settle. That's a different question from whether it'll do so to an inspecific prompt thrown at it.

I also don't want mood playlists, the majority of the time. I'm saying I use AI generations for exactly the opposite — so that I can explore and listen to things that are more intellectually interesting in the ways I find intellectually interesting, because the human music it's easy to find is common denominator music.

The way you describe AI ('continuous stream of auditory stimuli') is the way I'd describe Spotify. Sure, you could use AI to make a faux Spotify, but, like, why would you? The popular stuff already has saturating supply, and it will sound much better than an AI generation.

I think moritzwarhier's replies are highlighting that I didn't miss the point. Maybe I expressed it badly.

I was using ‘consumer preference’ here as a proxy for ‘the things that people tend to care about.’ The point I'm highlighting is that it's weird that when AI audio comes up, there's a reliable crowd of people discrediting it for missing a thing that clearly doesn't matter that much to most people in most contexts.

It's like, sure you can want things from music that are to your specific taste, but it's like coming into a post about, idk, a folk band and complaining that it's not metal. You're allowed to like your thing, but clearly most music is allowed not to be metal, why is this music specifically bad for not being metal?

And in this case the point I'm making is stronger, in that AI audio actually unlocks a lot of ability to listen things that are ‘interesting and creative’ but not widely available because of consumer preference, so it's actually more like showing up to a folk metal fusion band and saying the problem with this band is that it isn't metal.

This argument always seemed weird to me because it's obviously the case that ‘intent and creativity’ is at best a slim minority of consumer preference. If I want a simple dubstep line with random romantic words meaninglessly peppered in it, or the same 4 chord pop song as every other pop song, or you know, anything popular, that's easy. If I want something that's saying something interesting in the topics I find interesting, or doing something new with the medium that's also well executed and listenable to on repeat, finding that is actually just hard, and it's rarely the stuff with the dollars and overlapping interests to get the high production values.

I'm not going to claim AI audio isn't also awash with popular themes and tropes, or that it's a bastion of creativity. I'm also not going to claim that the deepest, really creative ideas aren't expressed in human written works. There are enough people to make truly exceptional songs and prompt many truly mindless AI generations. And there's also nothing wrong with most songs optimizing for personal preferences that are not that; I'm not trying to 'argue against' popular music.

But I am going to claim, for me, that it just hasn't been practical to saturate my tastes from public media, and that most of the reason I personally listen to AI music is that I want something that says or does something I think is creative, exploratory, or intellectually interesting that I don't know how to get from anywhere else.

The point being made is equivalent to pointing out that by strict mathematical definition there aren't physically reliazable distinguishably-non-Markovian processes, because computers are finite objects. Yes, you can talk about Markov chains as referring to everything up to and including actual human brains, but the kind of Markov chains people actually refer to as Markovian in practice are much more specific than that, typically those systems with structurally simple, indexical nodes, but sometimes extending to more complex ideas like RNNs where the states are at least of finite bandwidth. An LLM not only has continuous-valued nodes, it has nodes whose description length grows proportionally to and theoretically-unboundedly with respect to the size of its history.

The Rust example is illustrative, and indeed I was thinking about it when I wrote my post. One valid angle is to say that languages should hide the reality of the machine from the user any time it would get in the way of the pure semantic description. Another angle though is to say that actually these concerns are important, and if the result is that the idealised abstractions aren't sufficient to capture them, then it's correct to put away the abstraction. It's not like ‘simple and easy to learn’ is a function of this elegance either; Haskell is hard to learn and Go is easy, so clearly there's something else to it.

You also mention list comprehensions forming a monad. I think this is also a good illustration of the difference. In Haskell the structure of a type tells you the dependency tree of a computation, so it's not a problem that Monad maps the type to itself. Your list monad is just a bunch of thunks pointing to each other either way. In imperative languages, types describe what has been reified and how it's organised in memory. In these, a list monad is an actively bad abstraction; you almost always want to distinguish the stateful computational pipeline (eg. an iterator) from the source and target storage (eg. an array). Neither of those are monadic for good reason.

> and they allow you to support a whole family of things that look like "Promise<T>" and async functions, all using the same framework.

You've highlighted here the part that would actually explain the purpose of a monad, but not explained it. You don't need a monad abstraction to have things with monadic properties, and indeed often reality isn't quite perfectly shaped to theory, so forcing your object to fit the abstraction can be costly. One very obvious cost is that you no longer get descriptive names of what the monadic bind does; you have to infer it from what you know of the type.

The one thing a monad abstraction definitely gives you is the ability to write code that's generic over all monads. This is weird because this almost never happens outside of strongly functional languages.

If you'll forgive linking to a decade-old Reddit post, I've talked about this before.

https://reddit.com/r/programming/comments/4t6a6q/functional_...

(also, parent for context: https://reddit.com/r/programming/comments/4t6a6q/comment/d5f...)

The energy use claims are questionable, but I at least get where they're coming from. The water use is the confusing part. Who looks at a server rack and goes ‘darn, look at how water intensive this is’? People use water as a coolant in large part because it's really hard to boil, plus it's typically cheap because it regularly gets delivered to your front door for free.

As to actual numbers, they're not that hard to crunch, but we have a few good sources that have done so for us.

Simple first-principles estimate: https://epoch.ai/gradient-updates/how-much-energy-does-chatg...

Google report: https://arxiv.org/abs/2508.15734

Altman claim inside a blog post: https://blog.samaltman.com/the-gentle-singularity

Gentoo should be able to be ideological without being stupid.

I get why water use is the sort of nonsense that spreads around mainstream social media, but it baffles me how a whole council of nerds would pass a vote on a policy that includes that line.

Yes, fair.